Multilevel-prompt foundation model for nematode instance segmentation

Abstract Objective: Accurate nematode instance segmentation is critical in biomedical informatics, enabling rapid detection and localization of nematodes to provide data support for biomedical research on human aging mechanisms, drug screening, and neurological disorders. Despite the impressive performance of visual foundation models in general image segmentation tasks, their accuracy declines with nematodes due to the organisms’ complex characteristics, including multi-scale variations, curled morphologies, and frequent occlusions. Therefore, we propose a multilevel-prompt Segment Anything Model (SAM) architecture, enhancing segmentation precision. Methods: this study leverages the SAM as a foundational framework and systematically investigates its few-shot learning performance under diverse prompting strategies across three public nematode datasets and a proprietary dataset from the China Manned Space Engineering program. The proposed novel architecture sequentially integrates global point prompts and local box prompts, offering enhanced multi-view guidance for segmenting multi-scale nematode instances. To evaluate the efficacy of our approach, we conduct comparative experiments against the original point-prompt SAM and fine-tuned single-level prompt SAM model. We also conducted comparative experiments with baseline models. Results: The results demonstrate the superior performance of the multilevel-prompt SAM, with substantially improvements in segmentation accuracy across all datasets. Using mAP 50 : 95 , mAP 50 , mAP 75 , Dice score and Hausdorff Distance as evaluation metrics, our approach notably achieved 15.21% and 3.51% higher mAP 50 : 95 on the Mating dataset compared to Mask2former and OMGSeg for small-scale nematode targets. Conclusion: This study validates the practical applicability of the proposed methods for analyzing experimental data from the biomedical experiments and the Chinese Space Station using computer vision technologies, highlighting its potential for advancing fundamental life science research.